Classification of myocardial blood flow based on dynamic contrast‐enhanced magnetic resonance imaging using hierarchical Bayesian models

• Published in: Journal of the Royal Statistical Society Series C: Applied Statistics Vol. 71; no. 5; pp. 1085 - 1115
• Main Authors: Yang, Yalei, Gao, Hao, Berry, Colin, Carrick, David, Radjenovic, Aleksandra, Husmeier, Dirk
• Format: Journal Article
• Language: English
• Published: Oxford Oxford University Press 01.11.2022
• Subjects: Algorithms; Bayesian analysis; Blood flow; Cardiovascular disease; Cardiovascular diseases; Classification; Coronary artery disease; dynamic contrast‐enhanced magnetic resonance imaging; Estimates; Etiology; Fields (mathematics); Heart diseases; hierarchical Bayesian model; Lesions; Magnetic resonance imaging; Markov analysis; Markov chains; Markov random fields prior; Measurement; Medical diagnosis; Medical imaging; myocardial blood flow; myocardial perfusion; Myocardium; Probabilistic models; Uncertainty
• ISSN: 0035-9254; 1467-9876
• DOI: 10.1111/rssc.12568

Dynamic contrast‐enhanced magnetic resonance imaging (DCE‐MRI) is a promising approach to assess microvascular blood flow (perfusion) within the myocardium, and the Fermi microvascular perfusion model is widely applied to extract estimates of the myocardial blood flow (MBF) from DCE‐MRI data sets. The classification of myocardial tissues into normal (healthy) and hypoperfused (lesion) regions provides new opportunities for the diagnosis of coronary heart disease and for advancing our understanding of the aetiology of this highly prevalent disease. In the present paper, the Fermi model is combined with a hierarchical Bayesian model (HBM) and a Markov random fields prior to automate this classification. The proposed model exploits spatial context information to smooth the MBF estimates while sharpening the edges between lesions and healthy tissues. The model parameters are approximately sampled from the posterior distribution with Markov chain Monte Carlo (MCMC), and we demonstrate that this enables robust classification of myocardial tissue elements based on estimated MBF, along with sound uncertainty quantification. A well‐established traditional method, based on a Gaussian mixture model (GMM) trained with the expectation–maximisation algorithm, is used as a benchmark for comparison.